Indus script: seal catalogues and text-structure analysis

Machine-readable catalogues of the Mohenjo-daro seals, text features coded from the excavation plates, and a reproducible pipeline of structural analyses - approached statistically, without assuming any language.

No decipherment is offered here, and none is claimed to be reachable from this evidence. The analytical results are largely negative and are documented as such.

What is in it

Mackay (1938) seal table	709 rows: type, size, material, block/house/room, level, field number. Hand-transcribed; validated against Mackay's own totals
Marshall (1931) seal table	560 rows: type, size, material, excavation area, depth
Plate codings	531 seals from all thirteen seal plates, 419 joined to find-spots; blind re-coding gives kappa 0.89
Constraint set	90 distributional rules, mined on half the corpus and scored on the held-out half (1.08% violations)
Comparisons	Sumerian seal legends and administrative text, Egyptian (Old/Middle and Later), Linear A/B, proto-cuneiform, Proto-Elamite - genre-, size- and length-matched

Documentation

Primary document: docs/RESEARCH_NOTE.md - the full preprint: background, data, method, results and limitations in one place. See also docs/ABSTRACT.md for a short abstract.

file	what it is for
docs/RESEARCH_NOTE.md	the preprint: full method, results and limitations
docs/SUMMARY.md	the findings in one page
docs/FINDINGS.md	all 41 numbered findings, with figures and caveats
docs/STATUS.md	every finding marked standing / qualified / retired / replicates published work
docs/APPROACH.md	how the tests are built and why; the controls used
docs/PRIOR_WORK.md	what replicates published work and what does not
docs/DATASETS.md	data dictionary for every file
docs/REPRODUCING.md	how to run it; what needs manual downloads
docs/LIMITATIONS.md	what bounds the conclusions
docs/GLOSSARY.md	terms and sign names
docs/ROADMAP.md	what is still open
docs/DATA_NOTICE.md	licences of the third-party corpora

Running it

./setup.sh     # clone the pinned corpora
./run_all.sh   # run the pipeline (~1-2 min without optional data)

Steps whose optional inputs are absent are skipped with a message. See docs/REPRODUCING.md.

Citation

Koshy, N. S. (2026). Indus script: machine-readable seal catalogues from Mackay (1938) and Marshall (1931), with photo-coded text features and a reproducible analysis pipeline. Zenodo. https://doi.org/10.5281/zenodo.22852769

Companion preprint (full method, results and limitations): https://doi.org/10.5281/zenodo.22852774

See also CITATION.cff.

Licences

• Code: MIT (LICENSE).
• Transcribed and coded data: CC BY 4.0 (LICENSE-DATA). These are our transcriptions of public-domain excavation reports.
• Third-party corpora: not redistributed; their own terms apply. See docs/DATA_NOTICE.md.

A note on how this repo is kept

Results that were later corrected, qualified or retired are marked in place rather than deleted. A reader following the numbers will see the type F trend fail to replicate, the "3 + arrow" formula reinterpreted, and a periodicity claim shrink after a length control. Those corrections are part of the evidence.

Pipeline order (run_all.sh)

extract > features > cluster > finalize (visual families) > motif, motif_dedup, site, objtype (sign/animal/site/object tests) > functional, crosscheck, g400, finalize2 (behaviour classes) > abroad (foreign finds) > mj_segment, mj_subst (Mohenjo-daro template) > m77 (replication) > numerals, stock, size_tab, robust, mj_motif > refine, rename_families (shape refinement) > export_ml > ml_sequence > align_cisi, align_m77 > neural_nextsign > mj_core > count_arrow > build_mackay_table > mackay_context, mackay_spatial, match_feasibility, phase_test, neighbourhood_test, numeral_test, object_parts, recode_agreement, build_marshall_table, marshall_test, constraint_miner, affix_test, slot_tests, slot_followups, conditioning_compare, conditioning_shape, length_control, two_register, astral_test

[if CDLI present] proto_elamite, seal_legends, name_slot, numbered_titles, typology [if Mackay text present] parse_mackay, findspots

• mwenge/lineara.xyz @ 43fe7cf and mwenge/linearb.xyz @ 84e0b00 (code dedicated to the public domain; inscription transcriptions from GORILA / DAMOS-type sources): only LinearAInscriptions.js and LinearBInscriptions.js are fetched (sparse).
• CISI digitisation (mayig/indus-valley-script-corpus, via indus_decipher): cisi_real_corpus.csv, 179 Mohenjo-daro seals with Parpola sign numbers.

Optional data: Egyptian comparison corpus

Thesaurus Linguae Aegyptiae, CC-BY-SA 4.0, as JSONL at data/tla_egyptian.jsonl (columns: hieroglyphs, transliteration, lemmatization, UPOS, glossing, translation, dateNotBefore/After). egyptian_compare.py runs the structural measures on it. Egyptian matters here because it is a MIXED system (phonetic signs, logograms, determinatives) with a large inventory, and because its part-of-speech tags allow the measures to be checked against grammatical category on a language where the answer is known.

Optional data: Marshall (1931) vol. II

Marshall's own "Tabulation of Seals" (printed pp. 402-405 = PDF pages 56-59) covers the 1922-27 seasons, i.e. seals independent of Mackay's. Save the volume as data/marshall1931_vol2.pdf; source (out of copyright): https://archive.org/details/in.ernet.dli.2015.107493 transcriptions/marshall1931_seal_table/ holds the hand transcription (COMPLETE: 560 rows, Nos. 1-557 plus three suffixed), merged by build_marshall_table.py into outputs/marshall_seal_table_transcribed.csv. parse_marshall_table.py also OCRs the table into outputs/marshall_seal_table.csv, PROVISIONAL: the OCR recovers 356 of roughly 450 rows, plate numbers for half of them, levels for 239 and sizes for 67. The table is ordered by seal type, so types can be recovered from block boundaries. A visual transcription pass (about ten page-column views) would make it solid, as was done for Mackay. Note: Marshall gives level BELOW THE MODERN SURFACE and only the excavation area (HR, VS, DK, C, E, SD...), not block/house/room, so it is not directly comparable with Mackay's datum levels.

Optional data: Mackay (1938) find-spots

• Save the OCR text of Further Excavations at Mohenjo-daro, vol. I (out of copyright) as data/mackay1938_vol1.txt. Source (Internet Archive full text): https://archive.org/stream/dli.jZY9lup2kZl6TuXGlZQdjZM9k0Iy.TVA_BOK_0009130/TVA_BOK_0009130_Further_excavations_at_mohenjo-daro_djvu.txt
• parse_mackay.py extracts seal number, field number, area, block/street, level and an estimated phase to outputs/findspots_mackay.csv (the context column lets you check each row by eye). Test: python3 tests/test_parse_mackay.py.
• Joining find-spots to texts needs data/concordance_cisi_mackay.csv (template in tests/concordance_template.csv): Mackay's numbers are not CISI numbers. The mapping is printed in CISI vol. 1 (Joshi & Parpola 1987); it has to be transcribed by hand.

Mackay (1938) seal table (in repo)

transcriptions/mackay1938_seal_table/ holds a hand transcription of Mackay's "Tabulation of Seals" (number, type, size, material, block/house/room or street, level, field number) - complete, 709 rows. build_mackay_table.py merges and checks it into outputs/mackay_seal_table.csv and adds area, DK-datum level and an estimated phase. See the NOTES file there for conventions and progress.